Skill issues?
/use-skills
It's all just context. 
{
  "model": "claude-opus-4-7",
  "messages": [...]          ← text
  "system": "..."            ← also text
  "tools": [...]             ← also text (JSON schemas)
  "temperature": 1.0         ← sampling parameter
  "top_p": 0.95              ← sampling parameter
  "top_k": 40                ← sampling parameter
  "max_tokens": 4096         ← sampling parameter
  "stop_sequences": [...]    ← sampling parameter
}
quality
  ▲
  │      ╱─────╲
  │    ╱         ╲___
  │  ╱                ╲___
  │╱                       ╲___
  └────────────────────────────▶  context size
        sweet spot       too much
Less context = Better output
— Anthropic, Effective context engineering for AI agents
Context = Money
— GitHub Copilot pricing
2023
"Put it in the prompt."
Lands in: system
What broke: "lost in the middle" — instructions buried in long prompts silently stopped being followed.
2024
"Tools as JSON schemas."
Lands in: tools
What broke: too many tools → model picks the wrong one. Schemas count against context.
late 2024
"CLAUDE.md / .cursorrules / AGENTS.md"
Lands in: system (same slot!)
What broke: files grew unbounded, paid every turn, buried rules stopped firing.
2025
"MCP — connect everything!"
Lands in: tools (same slot!)
What broke: 134K tokens of tool defs before the user says hello.
Oct 2025
"Skills — load on demand."
Lands in: system (descriptions), messages (bodies)
What broke: descriptions are fuzzy triggers — they misfire either way. Bodies still drift.
2026 ?
What is a Skill? 
a11y-audit/
└── SKILL.md
───
name: a11y-audit
description: Use when the user asks to
  audit accessibility of a component or
  page. Do NOT use for visual design or
  general code review.
───

# Accessibility audit

## Workflow
1. Render the component or load the URL
2. Run scripts/run-axe.sh for violations
3. Run scripts/keyboard-trace.sh for focus order
4. Cross-check against references/wcag-2.2.md
5. Group issues by WCAG level (A / AA / AAA)
When does a skill enter the context?
Automatically.
The model sees the skill's name + description in the system prompt.
When the user's request matches, the full SKILL.md body gets loaded. 
user: audit a11y on the checkout modal
→ matches description of a11y-audit
→ load SKILL.md body
Explicitly.
Invoke the skill by name with a slash command.
Skips the routing — loads immediately. 
user: /a11y-audit Modal.tsx
→ load SKILL.md body
There's more.
a11y-audit/
├── SKILL.md
├── references/
│   ├── wcag-2.2.md
│   ├── aria-patterns.md
│   └── common-fixes.md
└── scripts/
    ├── run-axe.sh
    └── keyboard-trace.sh
More context?
always:     name + description    ~100 tokens
on match:   SKILL.md body         < 5K tokens
on demand:  references/*.md       only when SKILL.md says to
on demand:  scripts/*.sh          executed — code never enters context
Each level loads only when the previous one says to.
a11y-audit in action.
# startup
system:  a11y-audit — audit accessibility of a component or page
         ~100 tokens. that's all.

user:    audit a11y on the checkout Modal

claude:  cat a11y-audit/SKILL.md
         → instructions enter context

claude:  ./a11y-audit/scripts/run-axe.sh Modal
         → 4 violations (JSON). script code never enters context.

claude:  cat a11y-audit/references/wcag-2.2.md
         → cross-check violations

# aria-patterns.md, common-fixes.md — not needed here.
# never loaded. zero tokens.
Do skills actually work?
Yes*
Deep
Design tokens live in src/tokens/. Spacing uses a 4px base grid.
Never use raw pixel values — always var(--space-{n}).
The Overlay component traps focus and restores it on close;
new modals MUST use it, not div with z-index.
Shallow
Ensure proper ARIA labels, keyboard navigation, and focus management.
Prefer <button> over <div>. Test with keyboard-only navigation.
Sounds expert. It's still just MDN.
SkillsBench          curated, deep skills        +16.2pp
                     self-generated shallow      −1.3pp

Skills in the Wild   34k random skills           worse than none
Breadth without depth is noise.
Write few. Write deep.
Write what the model can't guess.

Resources